def k_fold(k, label, epochs, params, load_best_weigth, verbose, TensorB,
name_of_best_weight, base_model):
flag = params['agumentation']
data = ld.load_data(
label, phase="aug_evaluation") if flag == True else ld.load_data(
label, phase="evaluation")
results = []
size = len(data['x']) // k
tmp_idx = np.arange(data['x'].shape[0])
np.random.shuffle(tmp_idx)
x = data['x'][tmp_idx]
y = data['y'][tmp_idx]
np.save("x.npy", x)
np.save("y.npy", y)
acc_vec = []
for i in range(k):
x_test = x[i * size:(i + 1) * size]
y_test = y[i * size:(i + 1) * size]
x_train = np.append(x[0:i * size], x[(i + 1) * size:], axis=0)
y_train = np.append(y[0:i * size], y[(i + 1) * size:], axis=0)
tmp = random.sample(range(len(x_train)), 232)
x_val = []
y_val = []
for j in tmp:
x_val.append(x_train[j])
y_val.append(y_train[j])
x_val = np.array(x_val)
y_val = np.array(y_val)
x_train = np.delete(x_train, tmp, axis=0)
y_train = np.delete(y_train, tmp, axis=0)
##########fixing data#########
data = ld.fix_data(flag, x_train, y_train, x_val, y_val, x_test,
y_test)
model = DTL(params=params,
base_model=base_model,
label=label,
data=data)
model.train(epochs, load_best_weigth, verbose, TensorB,
name_of_best_weight + str(i) + ".h5", "k_fold")
results.append(model.evaluate())
print(results[-1])
model.clear()
del model
return results
def k_fold(k, epochs, params1, params2, load_best_weigth, verbose, TensorB,
name_of_best_weight, base_model):
data_tmp = [None, None, None]
flag = None
if params2['agumentation']:
data_tmp[0] = ld.load_data('a', phase="aug_evaluation")
data_tmp[1] = ld.load_data('v', phase="aug_evaluation")
data_tmp[2] = ld.load_data('h', phase="aug_evaluation")
flag = True
else:
data_tmp[0] = ld.load_data('a', phase="evaluation")
data_tmp[1] = ld.load_data('v', phase="evaluation")
data_tmp[2] = ld.load_data('h', phase="evaluation")
flag = False
data = {}
data['x'] = data_tmp[0]['x']
data['y'] = [data_tmp[0]['y'], data_tmp[1]['y'], data_tmp[2]['y']]
size = len(data['x']) // k
tmp_idx = np.arange(data['x'].shape[0])
np.random.shuffle(tmp_idx)
x = data['x'][tmp_idx]
y = [
data['y'][0][tmp_idx], data['y'][1][tmp_idx], data['y'][2][tmp_idx]
]
acc_vec = []
np.save("x.npy", x)
np.save("y.npy", y)
for i in range(k):
x_test = x[i * size:(i + 1) * size]
y_test = [
y[0][i * size:(i + 1) * size], y[1][i * size:(i + 1) * size],
y[2][i * size:(i + 1) * size]
]
x_train = np.append(x[0:i * size], x[(i + 1) * size:], axis=0)
y_train = [
np.append(y[0][0:i * size], y[0][(i + 1) * size:], axis=0),
np.append(y[1][0:i * size], y[1][(i + 1) * size:], axis=0),
np.append(y[2][0:i * size], y[2][(i + 1) * size:], axis=0)
]
y_test1 = np.array([[y[0][i][0], y[1][i][0], y[2][i][0]]
for i in range(i * size, (i + 1) * size)])
y_train1 = [[y[0][i][0], y[1][i][0], y[2][i][0]]
for i in range(i * size)
] + [[y[0][i][0], y[1][i][0], y[2][i][0]]
for i in range((i + 1) * size, len(y[2]))]
y_train1 = np.array(y_train1)
tmp = random.sample(range(len(x_train)), 232)
x_val = []
y_val = [[], [], []]
y_val1 = []
for j in tmp:
x_val.append(x_train[j])
y_val[0].append(y_train[0][j])
y_val[1].append(y_train[1][j])
y_val[2].append(y_train[2][j])
y_val1.append(
[y_train[0][j][0], y_train[1][j][0], y_train[2][j][0]])
y_val1 = np.array(y_val1)
x_val = np.array(x_val)
x_train = np.delete(x_train, tmp, axis=0)
y_train[0] = np.delete(y_train[0], tmp, axis=0)
y_train[1] = np.delete(y_train[1], tmp, axis=0)
y_train[2] = np.delete(y_train[2], tmp, axis=0)
y_train1 = np.delete(y_train1, tmp, axis=0)
for j in range(3):
y_val[j] = np.array(y_val[j])
##########fixing data#########
data2 = ld.fix_data(flag, x_train, y_train, x_val, y_val, x_test,
y_test)
data1 = ld.fix_data(flag, x_train, y_train1, x_val, y_val1, x_test,
y_test1)
model = DMTL(params=params1,
base_model=base_model,
label="",
loss='binary_crossentropy',
second_model=None,
phase=1,
data=data1)
model.train(1,
False,
verbose,
TensorB,
name_of_best_weight + str(i) + ".h5",
"k_fold",
save_m=True)
epochs = [1, 1, 1]
labels = ['a', 'v', 'h']
for l in range(3):
second_model = load_model("model_" + name_of_best_weight +
str(i) + ".h5",
custom_objects={
"a": Meric.a,
"v": Meric.v,
"h": Meric.h,
"loss_a": Meric.loss_a,
"loss_v": Meric.loss_v,
"loss_h": Meric.loss_h
})
model = DMTL(params=params2,
base_model=base_model,
label=labels[l],
loss='binary_crossentropy',
second_model=second_model,
phase=2,
data=data2)
model.train(epochs[l],
load_best_weigth,
verbose,
TensorB,
name_of_best_weight + str(i) + ".h5",
"k_fold",
save_m=False)
ans = model.evaluate()
acc_vec.append(ans)
print(ans)
model.clear()
del model
return acc_vec